feat: LLM并发模式统一——batch_reassess/trade_capture加concurrent=True(盘前批量4worker不再全压同一key,router round-robin分key), parallel_batch worker 4→6(6 key用满)

deploy/profile-scripts/batch_reassess.py

@@ -998,7 +998,7 @@ def process_stock(code, force_today=False):
     prompt = build_prompt(data)
     
     # ── 使用共享 LLM 客户端（替代 curl subprocess）──
-    result = call_llm(prompt, model=REASSESS_MODEL, max_tokens=None)  # 文档: 推理模型不指定max_tokens
+    result = call_llm(prompt, model=REASSESS_MODEL, max_tokens=None, concurrent=True)  # 2026-08-24 并发模式: router round-robin分key(4worker不再全压同一key)
     
     if not result["ok"] or not (result.get("content") or "").strip():
         print(f"  \u274c LLM调用失败或空输出: {result.get('error') or 'empty content'}")
@@ -1012,7 +1012,7 @@ def process_stock(code, force_today=False):
     # ── 截断保护：输出过短且无信号 = 低质输出，升级 pro 重试一次 ──
     if not parsed.get("signal") and len(full_text) < 1500:
         print(f"  ⚠️ 输出截断({len(full_text)}字)且无信号，升级 {FALLBACK_MODEL} 重试...", flush=True)
-        result2 = call_llm(prompt, model=FALLBACK_MODEL, max_tokens=None)  # 文档: 推理模型不指定max_tokens
+        result2 = call_llm(prompt, model=FALLBACK_MODEL, max_tokens=None, concurrent=True)  # 2026-08-24 并发模式
         if result2["ok"] and len((result2.get("content") or "").strip()) > len(full_text):
             full_text = result2["content"]
             parsed = parse_response(full_text)

deploy/profile-scripts/parallel_batch.sh

@@ -2,7 +2,7 @@
 # parallel_batch.sh — 并发批量12维重评（分片+key池偏移+统一flush）
 # 用法: bash parallel_batch.sh [N] [type]
 #   N=worker数(默认4)  type=holding|watchlist|all(默认all)
-N=${1:-4}
+N=${1:-6}  # 2026-08-24 4→6: OCG router 6 key 用满(concurrent=True round-robin防撞)
 TYPE=${2:-all}
 cd /home/hmo/MoFin/deploy/profile-scripts
 echo "[parallel] launching $N workers (type=$TYPE) at $(date '+%F %T')"

deploy/profile-scripts/trade_capture.py

@@ -179,7 +179,7 @@ def process_screenshot(ocr_text):
 【成本价】数字（有持仓时填写）
 """
     
-    result = call_llm(prompt, model=REASSESS_MODEL, max_tokens=None)
+    result = call_llm(prompt, model=REASSESS_MODEL, concurrent=True, max_tokens=None)
     if not result["ok"] or not result.get("content"):
         return {"ok": False, "error": "LLM调用失败"}